The artificial intelligence (AI) revolution is upon us, transforming not just our daily lives with smart assistants, personalized recommendations, and autonomous systems but also profoundly altering the landscape of scientific research and knowledge discovery. This revolution is characterized by the integration of AI into every domain of human activity, from healthcare and finance to education and entertainment. Its transformative effects are also being felt in the world of scientific research, e.g., in the physical sciences, where AI is not just assisting in data analysis1 but is also driving new discoveries2 and pushing the boundaries of knowledge and applied sciences. The interplay between AI and physics has reached a point where advancements in one field are catalyzing progress in the other, creating feedback loops of innovations that are reshaping our understanding of the universe and the tools we use to explore it.

The interaction between physics and AI has been a symbiotic one, where principles from physics have been applied to enhance AI models, and AI, in turn, has been used to solve complex problems in physics. This dynamic interplay is beautifully exemplified by the work of the 2024 Nobel laureates in Physics, John Hopfield3 and Geoffrey Hinton.4 Some of their pioneering contributions to artificial neural networks,5^–16 which are deeply rooted in concepts borrowed from or inspired by physics,6^,7^,11^,12 have laid the foundations for the modern AI revolution; see Fig. 1. John Hopfield, a physicist by training, was one of the first to draw a strong connection between physics and neural networks. His work was inspired by the complex world of spin glasses—disordered magnetic systems with intricate interactions. Hopfield recognized an analogy between these physical systems and networks of interconnected neurons in the brain.7 This insight led him to develop a type of recurrent neural network, known as the Hopfield network,17 capable of storing and retrieving patterns [Fig. 1(a)]. His 1982 paper,7 a cornerstone in the field of AI, demonstrated how principles from condensed matter physics could be harnessed to create computational systems that have learning and memory. The Hopfield network provided methods to explore how associative memory18 works, both in biological systems and artificial ones, and became one of the foundational models for AI research. Hopfield’s work was groundbreaking not just because it connected physics to AI but because it introduced the concept of energy landscapes9^,19 to neural networks. In a Hopfield network, the system settles into states of minimum energy, akin to how a physical system seeks equilibrium. This analogy to physical systems allowed researchers to use well-established methods from statistical mechanics to analyze and better understand neural networks, which opened up new avenues for creating advanced AI systems that could more closely emulate human cognition.

Similar to Hopfield’s scientific explorations at the intersection of physics and AI, Geoffrey Hinton, a cognitive psychologist and computer scientist, also took the push–pull relationship between AI and physics as a major inspiration for his seminal work. Hinton recognized the potential of Boltzmann machines,12 a type of stochastic neural network inspired by statistical mechanics, to learn complex patterns from data [Fig. 1(b)]. One of Hinton’s groundbreaking contributions was developing efficient learning algorithms for these networks.14^,15^,20 These algorithms enabled neural networks to extract meaningful features from data, such as images, text, or language, by optimizing the network’s parameters through processes similar to energy minimization. He also popularized the backpropagation algorithm,13^,21^,22 revolutionized convolutional neural networks,23 and introduced techniques such as dropout to improve training.24 Hinton’s work laid the groundwork for modern deep learning architectures, the applications of which have revolutionized fields such as computer vision, natural language processing, robotics, and biomedical sciences.

Beyond the seminal works of Hopfield and Hinton, the influence of physics on AI extends to various other areas. For instance, the concept of renormalization group, a powerful tool in condensed-matter and particle physics used to study systems with many interacting components across different scales, has found applications in deep learning for analyzing hierarchical structures and improving the efficiency of training algorithms.25 Another example is the use of quantum mechanics principles to develop new types of neural networks, known as quantum neural networks,26 which leverage quantum phenomena such as superposition and entanglement to potentially achieve exponential speedups for certain computational tasks. Furthermore, ideas from information theory, a field with deep roots in thermodynamics and statistical mechanics, have been instrumental in developing algorithms for compressing and efficiently representing information in AI systems. As another important example, diffusion models,27 a new class of powerful generative models, draw direct inspiration from the physics of diffusion28^,29 and Brownian motion. These examples illustrate the rich and ongoing cross-fertilization between physics and AI, where fundamental concepts from physics continue to inspire novel approaches and solutions in the realm of AI.

The impact and uses of AI in physics extend far beyond data analyses and simulations. It is fostering a deeper understanding of fundamental principles, enabling the design of entirely new physical systems. This influence is particularly evident in fields, such as optics and photonics,30^–32 where AI is revolutionizing the way that scientists manipulate, control, and harness light. From designing novel optical materials with unprecedented properties to optimizing the performance of complex photonic devices, AI is pushing the boundaries of possibilities regarding the manipulation and control of light, unveiling exciting new possibilities for applications in computing, sensing, imaging, and beyond.

One of the exciting areas where AI has been making a significant impact is computational imaging and sensing. Traditional imaging methods often face limitations due to the physical constraints of optics, such as resolution limits or noise. AI, however, is offering powerful new tools to mitigate some of these barriers. In microscopy, for example, AI algorithms can enhance image resolution, remove noise and artifacts, and even reconstruct 3D structures from limited data.33 Techniques such as super-resolution microscopy, which breaks the diffraction limit of light to reveal finer details than previously possible, have been significantly advanced by AI.34^–40 In holographic imaging, AI algorithms have excelled in solving complex physics-based inverse problems,41^–46 such as reconstructing a 3D scene from holographic data,47 with greater accuracy and speed than traditional methods, also providing different contrast mechanisms, e.g., reconstructing the images of specimens with brightfield contrast using their monochrome holograms.47 In fact, AI has been driving major innovations through such cross-modality image transformations,36^,48 where the spatial and spectral information typically associated with one imaging modality is extracted from data acquired using a different modality. This capability is opening up exciting new possibilities for biomedical imaging49 and remote sensing,50 among others. A compelling example is virtual staining in digital pathology and microscopy.51^,52 Traditional histological staining of tissue involves applying chemical stains to biological tissue to highlight various features under a microscope. However, this staining process can be time-consuming, laborious, and costly and can also damage the samples. Deep neural networks can now routinely transform label-free images of specimens into virtually stained microscopic images that mimic the appearance of traditionally stained images, eliminating the need for the chemical staining processes.49^,53 This allows for faster, cheaper, and more efficient analysis of biological samples and has significant implications for histology as well as live-cell imaging, where minimizing or eliminating chemical perturbations to the native biological system (through, e.g., external labeling and tags) is crucial.

AI is also making significant inroads in optical sensing, impacting both the design of sensors and the interpretation of sensor data.54 In areas such as biosensing and environmental monitoring, AI algorithms can rapidly process complex optical signals to detect subtle changes and identify specific analytes or conditions with greater sensitivity and specificity.55^–63 AI is also being used to design novel optical sensors with improved performance. For instance, in the development of optical sensors for point-of-care diagnostics, AI can optimize the design of the optical detection systems to enhance sensitivity/specificity and reduce sample volume requirements, while also providing multiplexed detection that can be used for the rapid and quantitative measurement of a panel of biomarkers and disease conditions.63^–68 By automating the optimization, quantitative multiplexed sensing, and decision processes, AI is accelerating the development of innovative optical sensors with tailored functionalities for various applications in point-of-care sensing, diagnostics, and environmental monitoring, as well as structural health, among many others.54^,69^–72

AI is also revolutionizing the design of optical materials, devices, and systems73^–76 by enabling a paradigm shift in “inverse” design.30^,77 Traditional inverse design approaches in optics and photonics typically rely on iterative optimization algorithms. These algorithms start with an initial guess of the device structure and repeatedly simulate its performance, using the results to refine the design parameters. This process continues until the desired performance metric is achieved. While these methods can be effective, they often require significant computational resources and time, especially for complex tasks and designs. In contrast, deep learning-based approaches offer more efficient and powerful alternatives. These alternative methods employ training a neural network on a large dataset of optical structures and their corresponding performance metrics.78 Once trained, the network can rapidly predict the performance of new designs and even generate novel structures with desired properties. This learning-based approach significantly accelerates the design process and enables the exploration/optimization of a wider range of parameters and possibilities.79^–81 AI-powered inverse design has already led to the creation of materials and systems with unprecedented capabilities. These include unidirectional imagers,82 invisibility cloaks that can render objects invisible,83 and ultra-efficient light absorbers for enhanced solar energy harvesting,84 among many others. Furthermore, this AI-powered optimization framework allows for the smart design of free-form optics, enabling the creation of compact and lightweight optical systems with superior performance.85^–87

Despite the remarkable progress made at the intersection of physics and AI, its widespread adoption faces some bottlenecks.88^,89 Some of these challenges are primarily related to high energy consumption, bandwidth limitations, and latency of AI systems, as well as hallucinations/artifacts in inference. For example, while AI offers immense potential in computational imaging, hallucinations in the generated/reconstructed images create concerns, as there might be features or details in the output images that are not present in the original data, which could be catastrophic, especially for biomedical imaging-related applications. This can occur in both inverse problems and cross-modality image transformations, leading to inaccurate reconstructions or misleading interpretations. To mitigate some of these issues, researchers have been incorporating physics-based loss functions into the training of AI models. These loss functions penalize deviations from known physical principles, guiding the learning and inference of the AI model to generate outputs consistent with the underlying physics of the system.52^,90 For example, in holographic image reconstruction, a physics consistency-based loss function was used to incorporate knowledge about the wave equation, driven from Maxwell’s equations, ensuring that the reconstructed scene adheres to the laws of wave propagation in free-space. This was shown to prevent the generation of unrealistic artifacts and significantly improve the fidelity of the holographic image reconstructions for out-of-distribution objects, showing superior external generalization behavior based on physics consistency-driven learning.90

Another challenge for future AI systems is that training and rapidly running complex AI models require immense computational power, leading to substantial energy demands91 and large carbon footprints.92 In addition, transferring vast amounts of data between memory and processing units can strain bandwidth and introduce latency, making real-time applications potentially difficult to implement through very large-scale models.93 This is one of the areas where optics and photonics might offer promising solutions.30^,87^,89^,94^–97 Optical computing platforms leverage the inherent parallelism and speed of light to perform, e.g., matrix multiplications and other computationally intensive tasks with significantly lower energy consumption and latency compared to electronic systems.98^–104 Free-space optical computing platforms, such as diffractive optical networks and smart metasurfaces, can perform visual computing, i.e., directly executing analog computation on visual information from an input scene without the need for digitization or preprocessing of information, enabling massively parallel processing of optical information with minimal energy dissipation. These approaches offer a powerful platform for the implementation of frontend analog information processing, delivering a compressed form of representation to back-end digital neural networks,87^,94^,105^,106 providing us with the best of both worlds.107^–110 This collaboration between optical analog processing and digital processors using neural networks can also swap places, where neural networks are used as digital encoders of information for optical networks106^,111^–113 to decode with extreme parallelism, requiring no external power except for the illumination light.

As physics and AI continue to drive innovations in the optics and photonics field, the synergy between these two disciplines will inevitably deepen. Physics has been providing foundational principles that guide AI development, while AI has been helping to unravel complex physical phenomena, offering new advances in various fields, from quantum mechanics to cosmology. However, challenges remain in fully integrating AI into physical sciences. The “black box” nature of many AI algorithms can hinder physical interpretability and trust, making it difficult to comprehend the underlying principles driving AI-generated solutions. In addition, ensuring that AI models generalize accurately and avoid “hallucinations” or spurious results requires careful validation and a robust feedback loop between physics and deep learning systems. By fostering a deeper integration that addresses these challenges with proper regulations and checks and balances, we can create large-scale AI-powered models and systems that are not only innovative but also reliable, interpretable, and capable of pushing the boundaries of scientific discovery and technological advancements in physical sciences at large. The recent Nobel Prize in Physics, awarded for pioneering work at the intersection of artificial neural networks and physics, signals the tremendous potential of this push–pull relationship between AI and physics, heralding a new era of research where they propel each other towards groundbreaking advances and discoveries.

Biographies of the authors are not available.